<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom" xmlns:dc="http://purl.org/dc/elements/1.1/">
  <channel>
    <title>DEV Community: makenna!</title>
    <description>The latest articles on DEV Community by makenna! (@siamakenna).</description>
    <link>https://dev.to/siamakenna</link>
    <image>
      <url>https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4183435%2F94048933-9264-4eab-b471-a368af3740b4.jpg</url>
      <title>DEV Community: makenna!</title>
      <link>https://dev.to/siamakenna</link>
    </image>
    <atom:link rel="self" type="application/rss+xml" href="https://dev.to/feed/siamakenna"/>
    <language>en</language>
    <item>
      <title>I Built BenchSloth to Keep Track of What Still Applies</title>
      <dc:creator>makenna!</dc:creator>
      <pubDate>Sun, 25 Oct 2026 00:10:11 +0000</pubDate>
      <link>https://dev.to/siamakenna/my-ai-lost-to-simple-rules-so-i-built-benchsloth-around-evidence-5966</link>
      <guid>https://dev.to/siamakenna/my-ai-lost-to-simple-rules-so-i-built-benchsloth-around-evidence-5966</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week3-2026-10-19"&gt;Hacktoberfest Open-Source AI Challenge: Week 3&lt;/a&gt;: Slow and Steady.&lt;/em&gt;&lt;/p&gt;

&lt;p&gt;Part of my work happens at a computer, and part happens at the bench. I wanted a calmer way to move between them without losing the small instructions that determine what should happen next.&lt;/p&gt;

&lt;p&gt;I started by asking AI to find waiting time in a protocol. Longer instructions led me to a different problem: &lt;strong&gt;keeping the source, its exceptions and the reason for a decision connected when the protocol changes.&lt;/strong&gt;&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;BenchSloth is a local attention organizer for research notes. Original passages sit beside an Attention Map, a movable evidence pinboard and a review queue. It doesn't execute an experiment or observe physical progress.&lt;/p&gt;

&lt;p&gt;The desk can be resized, rearranged or reduced to one region in Focus Mode. Each pin keeps its original occurrence ID and text. Keyboard controls and compact mobile regions provide alternatives to dragging. Moving a card changes the workspace, not the protocol.&lt;/p&gt;

&lt;p&gt;The Revision Impact Desk connects that workspace across versions:&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Compare → inspect impact → review evidence → confirm or abstain → recover context → return to the source.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu3u6dn5idsnpeytc3dti.jpg" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fu3u6dn5idsnpeytc3dti.jpg" alt="BenchSloth Attention Desk: original source, attention map and evidence workspace" width="800" height="500"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://benchsloth-hf26-week3-2026.onrender.com" rel="noopener noreferrer"&gt;Try BenchSloth in your browser →&lt;/a&gt;&lt;/strong&gt; No account, model download or installation required. This is the v0.9.1 static demo; its workspace, decisions and countdowns run in your browser.&lt;/p&gt;

&lt;h2&gt;
  
  
  Keeping the reason beside the revision
&lt;/h2&gt;

&lt;p&gt;Here is a synthetic example, created independently of the private SOPs:&lt;/p&gt;

&lt;blockquote&gt;
&lt;p&gt;OLD: “Wait 7 minutes.”&lt;/p&gt;

&lt;p&gt;OLD exception: “Do not use source 2 if the seal is damaged.”&lt;/p&gt;

&lt;p&gt;NEW: “Wait 9 minutes.”&lt;/p&gt;

&lt;p&gt;NEW exception: “Do not use source 2 unless the seal has been inspected.”&lt;/p&gt;
&lt;/blockquote&gt;

&lt;p&gt;The number changes, but the exception matters too. I want to inspect both versions, their monitoring language and the reason I previously left the question open.&lt;/p&gt;

&lt;p&gt;BenchSloth flags literal duration, monitoring, order and deletion cues. It keeps OLD and NEW IDs separate. A link brings an exception into view; it does not establish a scientific dependency or prove that the two passages mean the same thing. Display pairings can remain uncertain.&lt;/p&gt;

&lt;p&gt;After inspecting the originals, I can confirm a correspondence, reject it or abstain. An optional rationale stays with the exact hashes, source IDs and displayed context. Editing either version makes the earlier annotation stale. Its original reason remains available, and Undo retains the decision history.&lt;/p&gt;

&lt;p&gt;Even identical wording does not carry approval forward: an exception elsewhere may have changed. Opening NEW is a separate confirmation, with no inherited source confirmations or timer permissions.&lt;/p&gt;

&lt;h2&gt;
  
  
  Tracking a duration without granting a break
&lt;/h2&gt;

&lt;p&gt;The earlier unattended-timer gate offered zero intervals across five real SOPs. That did not mean there was no useful timing information. Compound instructions, ranges, active work and equipment language often fell outside its narrow grammar.&lt;/p&gt;

&lt;p&gt;I separated &lt;strong&gt;tracking a source duration&lt;/strong&gt; from &lt;strong&gt;deciding whether something may be left unattended&lt;/strong&gt;.&lt;/p&gt;

&lt;p&gt;The tracker requires an exact original token, my interpretation of it, a review of the restrictions and a manual start. A range needs an explicit value within its bounds; there is no automatic midpoint. “Overnight” stays qualitative. Choosing an hour from “an hour or overnight” does not settle that alternative.&lt;/p&gt;

&lt;p&gt;Concurrent trackers retain separate source IDs. Retrieved original conditions stay beside the clock, with their applicability still for review. Replacing the source invalidates the older trackers.&lt;/p&gt;

&lt;p&gt;Breathing Room is quiet, but it does not imply that observation can stop. The demonstration counts down ten real seconds while preserving the source's seven-minute wording. Expiry does not confirm the step or complete the science. Browser clocks are a convenience, and reload does not resume an alarm.&lt;/p&gt;

&lt;h2&gt;
  
  
  Help me pick up where I left off
&lt;/h2&gt;

&lt;p&gt;Context Recovery brings together five things: where the browser workflow stopped, what changed, which conditions need attention, which questions remain open and where to look next.&lt;/p&gt;

&lt;p&gt;The complete handoff works without AI. It includes current source links, version-qualified comparisons, organizer annotations and tracker status. A later decision or source change marks it stale rather than quietly presenting old context as current.&lt;/p&gt;

&lt;p&gt;I implemented one final, bounded Gemma 3 4B candidate for explaining that context. It receives original excerpts, exact changes and server-recorded state through a typed JSON contract. Browser-only organizer annotations and trackers remain in the complete deterministic handoff; the server does not pretend to know them.&lt;/p&gt;

&lt;p&gt;A citation proves identity, not that a paraphrase follows from the evidence. Noncanonical model wording remains unverified. Wrong source versions, stale state, invalid references and missing handoff sections cannot become accepted instructions. No response can authorize a timer.&lt;/p&gt;

&lt;p&gt;In the final local verification, &lt;strong&gt;27 of 28 frozen scenarios were attempted&lt;/strong&gt;: 26 returned responses, one timed out at the unchanged 100-second limit, and the last case was left unattempted under the stop rule. Only &lt;strong&gt;3/27 attempted workflows&lt;/strong&gt; produced complete exact handoffs, versus &lt;strong&gt;27/27&lt;/strong&gt; for the paired deterministic baseline. Mean returned-request time was &lt;strong&gt;57.98 seconds&lt;/strong&gt;. A textual audit found one invented edit assertion; the application rejected it. Most other failures repeated sections or omitted context. The public &lt;a href="https://github.com/siamakenna/benchsloth-hf26-week3-2026/tree/main/evidence/gemma" rel="noopener noreferrer"&gt;evaluation summary and raw audit&lt;/a&gt; preserve those results; the complete local captures remain archived.&lt;/p&gt;

&lt;p&gt;The complete deterministic handoff remains the default. I have not established a communication or usefulness advantage for the model, and no independent person has scored readability or correction effort. I am not promoting automatic evidence selection, revision classification or this explanation layer as a reliable semantic authority.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;The first version of BenchSloth was mostly concerned with finding waiting periods in a protocol. Once I tested it against longer laboratory instructions, I realized that wasn't the hardest part. A protocol can tell you to wait for an hour, prepare something fifteen minutes before that interval ends, and keep a sample under specific conditions throughout. Recognizing the hour doesn't tell you how those instructions fit together. And if the protocol changes, you need to know which earlier decisions no longer apply.&lt;/p&gt;

&lt;p&gt;That became the more interesting problem for me: not just counting time, but keeping the original instructions, their exceptions, and the reasons behind a decision connected. I ended up giving deterministic code more responsibility than the language model, which wasn't what I'd expected when I started.&lt;/p&gt;

&lt;p&gt;The application uses Python, FastAPI and plain browser JavaScript. The deterministic document path supports &lt;strong&gt;512 nonempty occurrences and 100,000 characters&lt;/strong&gt;, independently of the frozen model parser's smaller budget. Raw text, its hash and original offsets are retained. Headings, warnings and repeated identical instructions each have their own global occurrence ID.&lt;/p&gt;

&lt;p&gt;Explicit references and ordinals are resolved against the complete document before candidate limits apply. Oversized model context is rejected rather than shortened invisibly. Heading recognition remains heuristic; native document styling is not fully interpreted.&lt;/p&gt;

&lt;p&gt;I reused small engineering ideas from my other projects: stable identity contracts, frozen comparisons, blank review provenance and revision-bound evidence gates. No research datasets or scientific conclusions were imported. Earlier reused MIT code and local video tooling remain credited. I did not add partner services merely to list integrations.&lt;/p&gt;

&lt;h2&gt;
  
  
  What the comparisons changed
&lt;/h2&gt;

&lt;p&gt;Gemma's failures changed the design more than its successful examples did.&lt;/p&gt;

&lt;p&gt;For source-grounded questions, a correct passage did not guarantee a correct relationship. The model could retrieve relevant evidence and still confuse “during” with “after.” On the earlier frozen 24-question comparison, the combined candidate produced &lt;strong&gt;13/24 fully correct workflows versus 14/24&lt;/strong&gt; for strong rules, with &lt;strong&gt;8 unsupported evidence selections versus 7&lt;/strong&gt;. Its mean local model-request time was &lt;strong&gt;25.40 seconds&lt;/strong&gt;, compared with about &lt;strong&gt;0.00067 seconds&lt;/strong&gt; for retrieval. These are software timings, not researcher time savings.&lt;/p&gt;

&lt;p&gt;Focused relationship classification and generated review briefs also lost to their templates. The revision pilot reached correspondence F1 &lt;strong&gt;0.889 versus 0.970&lt;/strong&gt; for the original deterministic baseline; exact semantic cards were &lt;strong&gt;7/20 versus 9/20&lt;/strong&gt;. A later deterministic correction reached &lt;strong&gt;11/20&lt;/strong&gt;, retained separately rather than substituted into the frozen result.&lt;/p&gt;

&lt;p&gt;The historical comparisons remain documented, and the public release includes the final Context Recovery evaluation contract, aggregate results and raw audit. The related synthetic examples and implementer-visible labels limit generalization. I have not rewritten the older results around a more flattering story.&lt;/p&gt;

&lt;h2&gt;
  
  
  What happened with the real SOPs
&lt;/h2&gt;

&lt;p&gt;The first private audit admitted only &lt;strong&gt;2 of 5&lt;/strong&gt; complete SOPs. Their 70 admitted occurrences were quoted exactly, but the larger documents were rejected. Nothing was silently truncated; practical coverage was still poor.&lt;/p&gt;

&lt;p&gt;After separating document capacity from model capacity, all five full texts passed: &lt;strong&gt;48, 22, 58, 57 and 115 occurrences — 300/300 retained and quoted exactly&lt;/strong&gt;. The final regression run repeated those lookups and version/state checks with model transport forbidden.&lt;/p&gt;

&lt;p&gt;The source-timing pass found &lt;strong&gt;34 exact numeric tokens across 34 occurrences&lt;/strong&gt;. Test actions exercised confirmation guards, retained restrictions, expiry without step completion, source replacement and Undo. They are software-mechanics checks, not proof that every token is a useful interval or that no relevant duration was missed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Human-reviewed scientific accuracy remains unvalidated.&lt;/strong&gt; Eighteen source-bound cases are ready for owner review, with blank fields to accept, reject, correct or leave unresolved. The earlier 300-card audit and 50-card priority set are preserved. No model or automated test filled a human review field.&lt;/p&gt;

&lt;p&gt;Private posters, presentations and notebooks describe biological observations. I did not establish the exact SOP-version, sample and run provenance needed to compare those outcomes with this software audit. They are not evidence of BenchSloth's performance, and their contents stay private.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://benchsloth-hf26-week3-2026.onrender.com/media/walkthrough.mp4" rel="noopener noreferrer"&gt;Watch the 72-second walkthrough&lt;/a&gt;&lt;/strong&gt; · &lt;a href="https://benchsloth-hf26-week3-2026.onrender.com/media/captions.vtt" rel="noopener noreferrer"&gt;Captions&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The walkthrough uses actual viewport captures of the local v0.9 application, including its real ten-second demonstration tracker. It is an edited sequence of captures, not continuous video of a new Gemma inference. The deployed v0.9.1 adds private visual notes to that workflow.&lt;/p&gt;

&lt;p&gt;For a first minute, inspect a passage, pin it and start the clearly labeled demonstration. Compare OLD and NEW, inspect the changed restriction, record an abstention and edit NEW again. The earlier decision becomes stale. Open NEW separately, recover context and follow the current source link back to its original wording.&lt;/p&gt;

&lt;p&gt;The self-contained replay needs no Ollama. Resizing, pinning, literal lookup, review, manual timing and deterministic recovery run in the browser with runtime network connections denied. Historical Gemma outputs are labeled as recorded. Final native response records are separate experimental evidence, never a prepared substitute for a fresh request.&lt;/p&gt;

&lt;p&gt;The final pass verifies &lt;strong&gt;163 Python tests and 60 JavaScript assertions&lt;/strong&gt;, the actual loopback API/browser workflow, desktop and 320px/390px layouts, keyboard controls, quiet motion, source isolation and playback. Quiet controls and reduced-motion rules were checked; this is not a full assistive-technology, physical-device or independent usability study.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code and Version Scope
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://github.com/siamakenna/benchsloth-hf26-week3-2026" rel="noopener noreferrer"&gt;Public source and release&lt;/a&gt; · &lt;a href="https://benchsloth-hf26-week3-2026.onrender.com/media/evaluation.html" rel="noopener noreferrer"&gt;Evaluation page&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;The public v0.9.1 release contains the self-contained browser demo, selected Python and JavaScript implementation files, synthetic examples and sanitized evaluation evidence. It preserves the earlier v0.6 repository history without exposing the private development history. The selected source files are inspectable implementation evidence; this smaller package is not a complete one-command local server distribution.&lt;/p&gt;

&lt;p&gt;Render serves only an allowlisted static directory. It does not expose FastAPI, Ollama, private research inputs or a live model endpoint. The strict build checks each published file against its expected SHA-256 hash. Historical Gemma outputs are labeled as recorded; no prepared response is presented as fresh inference.&lt;/p&gt;

&lt;h3&gt;
  
  
  A photograph can be a note without becoming a conclusion
&lt;/h3&gt;

&lt;p&gt;The new visual-notes panel attaches a camera or uploaded image to a selected original passage, with an observation and an explicit manual review status. It performs &lt;strong&gt;no AI image interpretation&lt;/strong&gt;. Photos stay in browser-tab memory until the user explicitly exports a private JSON file; the site does not upload them.&lt;/p&gt;

&lt;p&gt;The browser makes a resized, lossy JPEG preview and strips embedded metadata. That preview is not primary scientific image data. Unaltered instrument images and acquisition metadata must stay in appropriate research storage. Changing the source marks earlier visual notes for another review. This is a way to retain context, not a claim that BenchSloth can assess a biological result.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Open Innovation Matters
&lt;/h2&gt;

&lt;p&gt;The application and original artwork are MIT-licensed; Gemma weights have separate terms. Private protocols and reviewer exports are excluded from the demo and public code.&lt;/p&gt;

&lt;p&gt;Open development makes it possible to inspect a failed request and the baseline that beat it, then see how the design changed. No independent researcher has completed the usability trial, and I have not measured productivity gains.&lt;/p&gt;

&lt;p&gt;For me, &lt;em&gt;Slow and Steady&lt;/em&gt; means leaving enough context for a decision to be reconsidered. The useful part is what should not disappear when instructions change: the original evidence, the exception and the reason someone left a question open.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Overall:&lt;/strong&gt; the research-organizer workflow is the main entry: preserving evidence, exceptions and decisions so returning to interrupted work takes less reconstruction. Actual time savings still need a human trial.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Render:&lt;/strong&gt; Render hosts the public static front end and walkthrough. No private research-processing server is deployed.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma:&lt;/strong&gt; the project contains genuine local Gemma 3 4B integration and a frozen comparison of its Context Recovery responses. This is experimental use, with its failures disclosed. The deterministic handoff remains the default; a model advantage is not established.&lt;/p&gt;




&lt;p&gt;AI tools supported the development of BenchSloth and the preparation of this article. The source code, evaluations, and recorded model behavior are available to inspect.&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
    </item>
    <item>
      <title>I Built FundPilot AI with Google Gemma to Make Grant Requirements Easier to Understand</title>
      <dc:creator>makenna!</dc:creator>
      <pubDate>Fri, 16 Oct 2026 07:14:06 +0000</pubDate>
      <link>https://dev.to/siamakenna/i-built-fundpilot-ai-with-google-gemma-to-make-grant-requirements-easier-to-understand-54ib</link>
      <guid>https://dev.to/siamakenna/i-built-fundpilot-ai-with-google-gemma-to-make-grant-requirements-easier-to-understand-54ib</guid>
      <description>&lt;p&gt;&lt;em&gt;This is a submission for the &lt;a href="https://dev.to/challenges/hacktoberfest-week2-2026-10-12"&gt;Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle&lt;/a&gt;.&lt;/em&gt;&lt;br&gt;
you&lt;/p&gt;

&lt;h2&gt;
  
  
  What I Built
&lt;/h2&gt;

&lt;p&gt;Finding funding for something you're passionate about should be exciting. Whether it's a research grant, a conference travel award, or a fellowship that could open the door to your next opportunity, financial support can make an enormous difference. Unfortunately, finding an opportunity is often easier than understanding whether you actually qualify for it.&lt;/p&gt;

&lt;p&gt;As a researcher myself, I've experienced how frustrating this process can be. You might discover an award that seems perfect, only to spend hours navigating eligibility requirements, exceptions, and application instructions before realizing that one small condition could change everything. For students, independent researchers, and people without access to dedicated funding support, this process can be especially overwhelming.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;That's why I built FundPilot AI, an open source funding assistant powered by Google's Gemma that helps people understand funding requirements and prepare for applications with greater clarity.&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;I wanted to create something I would genuinely use myself. With FundPilot, users can paste the public requirements from a funding announcement, answer a few straightforward questions, and receive an organized overview of what appears to match their circumstances, what might be a problem, and what still needs clarification.&lt;/p&gt;

&lt;p&gt;One of my favorite features is the &lt;strong&gt;Application Readiness Dossier&lt;/strong&gt;. Rather than simply identifying something as uncertain or potentially conflicting, FundPilot helps users understand what to do next. The downloadable dossier brings together possible obstacles, application materials, important dates, supporting evidence, and suggested questions to ask the funding organization.&lt;/p&gt;

&lt;p&gt;For example, if a conference travel award requires applicants to be the first author of a presentation, FundPilot can recognize that requirement, compare it with the information provided, and highlight a potential conflict. Instead of leaving someone wondering what that means, it connects the finding to the original announcement and suggests what should be verified before applying.&lt;/p&gt;

&lt;p&gt;Accessibility has been central to how I've approached this project. I wanted the experience to feel approachable even for someone unfamiliar with programming or AI. The latest version features a cleaner results dashboard, questions suggested from the funding text, and expandable sections that make the information easier to explore without becoming overwhelming.&lt;/p&gt;

&lt;p&gt;FundPilot does not make official eligibility decisions. My goal is to help people understand opportunities, recognize what needs attention, and feel more prepared to take their next step.&lt;/p&gt;

&lt;p&gt;I also see an exciting possibility for this to become an affordable funding preparation service, particularly for individuals and smaller organizations without grants offices. My initial business idea is &lt;strong&gt;$9 per readiness report or $39 per month for a funding workspace&lt;/strong&gt;. These are proposed prices, not existing paid services. I want to validate whether FundPilot saves people time and provides enough value before introducing payments.&lt;/p&gt;

&lt;h2&gt;
  
  
  Demo
&lt;/h2&gt;

&lt;p&gt;&lt;a href="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpiudj99pd7h8px6ai1h1.png" class="article-body-image-wrapper"&gt;&lt;img src="https://media2.dev.to/dynamic/image/width=800%2Cheight=%2Cfit=scale-down%2Cgravity=auto%2Cformat=auto/https%3A%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Farticles%2Fpiudj99pd7h8px6ai1h1.png" alt="FundPilot AI displaying funding requirements and potential problems" width="800" height="556"&gt;&lt;/a&gt;&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://youtu.be/Pw5Tco6q6lY" rel="noopener noreferrer"&gt;Watch FundPilot AI in Action&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;This 96 second demonstration shows FundPilot running with Gemma and turning public funding requirements into practical next steps. It walks through matching requirements, potential conflicts, missing information, and the Application Readiness Dossier.&lt;/p&gt;

&lt;p&gt;The recording includes four genuine Gemma analyses and eight verified report downloads. Applicant information is fictional, and the selected public passages are examples rather than complete reviews of the funding programs.&lt;/p&gt;

&lt;p&gt;You can also explore an &lt;a href="https://github.com/siamakenna/fundpilot-ai-week2-2026/blob/main/evidence/usability/contradiction-dossier.md" rel="noopener noreferrer"&gt;example Application Readiness Dossier&lt;/a&gt; or visit the &lt;a href="https://github.com/siamakenna/fundpilot-ai-week2-2026/releases/tag/v0.3.2" rel="noopener noreferrer"&gt;latest release&lt;/a&gt;.&lt;/p&gt;

&lt;h2&gt;
  
  
  Code
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;&lt;a href="https://github.com/siamakenna/fundpilot-ai-week2-2026" rel="noopener noreferrer"&gt;FundPilot AI on GitHub&lt;/a&gt;&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;FundPilot is built using Python, Gradio, Ollama, and Google's Gemma 3. The application code is publicly available under the MIT license, while Gemma's model weights are downloaded separately under Google's terms.&lt;/p&gt;

&lt;p&gt;The repository includes the working application, tests, example reports, and documentation explaining how the system was developed and evaluated.&lt;/p&gt;

&lt;p&gt;FundPilot currently runs locally, so users need to install the application and Ollama. Making setup simpler and eventually offering a browser based experience are important future goals, especially because I want the tool to become accessible to people without technical backgrounds.&lt;/p&gt;

&lt;h2&gt;
  
  
  How I Built It
&lt;/h2&gt;

&lt;p&gt;I started with a simple question: could a small AI model help someone understand complicated funding requirements while still allowing them to verify the original information?&lt;/p&gt;

&lt;p&gt;I used &lt;strong&gt;Gemma 3 1B&lt;/strong&gt;, running locally through Ollama, to help identify requirements, exclusions, deadlines, and other important details in public funding announcements. Python then compares supported conditions with the information supplied by the user and organizes the results into an understandable report.&lt;/p&gt;

&lt;p&gt;One decision I made early was to preserve the original evidence. When FundPilot highlights a requirement, the supporting passage comes directly from the announcement rather than being written by the AI. This matters because an incorrect interpretation could influence whether someone decides to pursue an opportunity.&lt;/p&gt;

&lt;p&gt;During development, I discovered that understanding funding language is much more complicated than simply summarizing a document. Some announcements contain exceptions, alternative ways to qualify, or conditions that depend on information the applicant hasn't provided. Instead of hiding that uncertainty, I designed FundPilot to make it visible and suggest what needs further clarification.&lt;/p&gt;

&lt;p&gt;The application also includes a separate &lt;strong&gt;Rule baseline (NO AI)&lt;/strong&gt; option, allowing users to analyze the same text without relying on Gemma. My evaluations against selected public funding passages showed that Gemma still has room to improve and does not consistently outperform the simpler rules. The evaluation references also need independent human review. I experimented with a larger Gemma model, but its slower processing and mixed results didn't justify replacing the smaller version.&lt;/p&gt;

&lt;p&gt;As I continued developing FundPilot, I realized that identifying a requirement was only part of what users needed. They also needed help understanding what the information meant for them and what action to take afterward. That realization inspired the verification system, the clearer results overview, and the Application Readiness Dossier.&lt;/p&gt;

&lt;p&gt;The latest version has &lt;strong&gt;147 passing automated tests&lt;/strong&gt;, along with successful checks of navigation, downloads, input handling, and layouts on smaller screens. These checks establish important aspects of software reliability, although they don't guarantee that every AI interpretation is correct.&lt;/p&gt;

&lt;p&gt;For anyone interested in the technical details, I've included the &lt;a href="https://github.com/siamakenna/fundpilot-ai-week2-2026/blob/main/docs/REAL_WORLD_EVALUATION.md" rel="noopener noreferrer"&gt;evaluation results and limitations&lt;/a&gt; in the repository.&lt;/p&gt;

&lt;h2&gt;
  
  
  Why Does Open Innovation Matter?
&lt;/h2&gt;

&lt;p&gt;One of the things I love most about open source technology is that it allows people to build solutions around problems they experience firsthand.&lt;/p&gt;

&lt;p&gt;I didn't start FundPilot simply because I wanted to build another AI application. I started it because I know how much funding opportunities can matter and how discouraging it can feel when understanding those opportunities becomes another obstacle.&lt;/p&gt;

&lt;p&gt;Using Gemma locally gave me the freedom to experiment, understand where the model made mistakes, and continue improving the application without paying for every analysis through a commercial AI service. Public announcement text is processed locally, while the applicant's answers remain within the Python application. This gives me more control over the process, although it doesn't mean the model is automatically more accurate or that running it has no hardware costs.&lt;/p&gt;

&lt;p&gt;Making FundPilot open source also allows others to examine the approach, learn from its development, and potentially contribute improvements. I hope it can eventually become a useful resource for researchers, students, nonprofits, and other people navigating funding applications without extensive institutional support.&lt;/p&gt;

&lt;p&gt;&lt;strong&gt;As someone whose own research journey has depended on grants, fellowships, and conference opportunities, this project feels especially personal.&lt;/strong&gt; I know how exciting it can be to find an opportunity that could move your work forward, and I want more people to experience that excitement without feeling overwhelmed by the application process.&lt;/p&gt;

&lt;p&gt;I don't believe AI needs to provide every answer to be valuable. Sometimes, helping someone identify an important requirement or understand the right question to ask is just as meaningful.&lt;/p&gt;

&lt;p&gt;My hope is that FundPilot can make funding information easier to understand, preparation less intimidating, and opportunities more accessible to the people pursuing them.&lt;/p&gt;

&lt;h2&gt;
  
  
  My First User
&lt;/h2&gt;

&lt;p&gt;One of the most encouraging parts of developing FundPilot was receiving early feedback from someone outside the development process. They appreciated the clean visual design and recognized the verification system as one of the project's strongest features.&lt;/p&gt;

&lt;p&gt;Their observations helped guide my latest improvements. I focused on making the experience easier to navigate through a simpler questionnaire, a clearer results dashboard, and supporting evidence that users can explore when they need more detail.&lt;/p&gt;

&lt;p&gt;This was an early qualitative review rather than a completed independent usability study. The participant did not pay for the product, and I have not yet measured time savings. I'm excited to continue gathering feedback and learning whether FundPilot helps people approach funding opportunities more efficiently.&lt;/p&gt;

&lt;h2&gt;
  
  
  Prize Categories
&lt;/h2&gt;

&lt;p&gt;&lt;strong&gt;Best Use of Gemma&lt;/strong&gt;&lt;/p&gt;

&lt;p&gt;FundPilot uses Google's Gemma 3 to help interpret public funding requirements and turn complicated information into a clearer, more accessible application preparation experience!&lt;/p&gt;

</description>
      <category>devchallenge</category>
      <category>hf26challenge</category>
      <category>ai</category>
      <category>career</category>
    </item>
  </channel>
</rss>
